Unsupervised Pre-Estimation GPS Integrity Screening for Connected Vehicle Localization
MSc Thesis Defense by: Nizbath Ahsan
Date: Tuesday, September 22nd, 2026
Time: 1:00pm to 2:30pm
Location: Essex Hall 122
Abstract:
Reliable localization is essential to connected-vehicle operation, because position and motion information feed vehicle state estimation and may support cooperative safety applications. If integrity is not assessed beforehand, manipulated GPS measurements can be accepted as legitimate observations by the estimation pipeline. This thesis investigates an unsupervised pre-estimation GPS integrity-screening framework that treats localization security as a measurement-admission problem. Isolation Forest and One-Class Support Vector Machine are trained only on benign driving data and used to classify incoming localization observations as normal or anomalous before they become eligible for downstream state estimation. The framework is evaluated in three heterogeneous localization settings derived from KITTI, highD, and DAIR-V2X, differing in driving environment, sensing availability, feature representation, attack prevalence, and decision calibration. Manipulation behaviours include spoofing, drift, freeze, replay-related or offset-based manipulation, and noise injection. Results show that benign-only anomaly detection can identify manipulated localization observations without labelled attack examples during training. OCSVM performs more strongly in the retained overall summaries and in many attack-specific conditions, although neither detector is uniformly superior, and performance varies substantially across attack types and localization representations. The highD experiments show that high attack recall can coexist with very low precision, making false-alarm control a critical requirement for practical measurement admission. Overall, this thesis establishes the feasibility of pre-estimation localization integrity screening using benign-only anomaly detection, and shows that screening effectiveness depends on the anomaly-detection algorithm, the available localization evidence, attack behaviour, and detector operating configuration. The evaluation is limited to the screening decision itself; downstream state-estimation accuracy and application-level vehicle behaviour are not assessed.
Keywords: Unsupervised anomaly detection, Isolation Forest, One-Class SVM, Localization integrity
Thesis Committee:
Internal Reader #1: Dr. Muhammad Asaduzzaman
Internal Reader #2 : Dr. Dan Wu
Advisor: Dr. Ikjot Saini
Chair: Dr Pooya Moradian Zadeh
